Machine Learning helps to separate entanglements in Bin-Picking Applications. The goal is to create a system that finds a path to separate an entanglement, starting from a single visual input. To realize such a system both supervised and reinforcement learning methods can be implemented. For both of these approaches we set up a motion model and the remaining properties of the real robot cell are implemented in a simulation scene. While the simulation scene can be used to create training data for the supervised learning approach, it is also the learning environment for the reinforcement learning model. Therefore, there are similar premises for comparing the two models. What needs to be investigated is which of the two methods separates the most entanglements and offers the least setup effort. The setup effort in general and the performance are examined for both approaches in simulation and also in real-world experiments. The reinforcement learning model outperforms both of the supervised learning models in the setup effort and the separation rate by over 15 percent points.